Blog

Dario Amodei's recent essay, The Adolescence of Technology, is one of the better articulations I've seen of where we actually are with AI. Not five decades out. Not sci-fi. More like an awkward teenage phase where the technology is getting powerful very fast, and the adults in the room are still arguing about curfews.

His core idea is simple but uncomfortable. We are about to give ourselves tools with leverage that looks historically unprecedented, and our institutions are not obviously ready to handle that responsibly. That sounds alarmist until you realize we already do this all the time. Social media scaled faster than governance. Financial engineering scaled faster than regulation. We usually clean things up after the damage is done.

What I appreciate is that Amodei is not anti-progress. He is aggressively pro-capability. The optimistic case is real. AI can unlock enormous productivity, accelerate science, improve healthcare, and grow the economic pie in ways that are hard to overstate. I work with these systems daily. They already feel like leverage, not novelty.

But optimism without guardrails is how you end up with bio tools in the wrong hands, surveillance states on steroids, or labor markets that break faster than society can adapt. His framing of a "country of geniuses" showing up overnight is intentionally jarring, and it should be. If that literally happened, we would not shrug and hope incentives work out.

So what needs to change, at least in the US? We need to treat AI as national infrastructure, not just a consumer product category. Transparency from frontier labs should be mandatory. Export controls on the hardest bottlenecks should be enforced. Civil liberties need explicit protection before AI makes abuse cheap and scalable. And economically, we need to get serious about redistribution mechanisms before displacement forces the issue for us.

The good news is that adolescence ends. If we do this right, AI becomes less a runaway experiment and more a durable advantage. If we do it wrong, we get a very expensive lesson in hindsight. History suggests we should probably try the first option this time.

The shift from manual coding to agent-driven development has been a total phase shift, and Andrej Karpathy's recent reflections on using Claude perfectly capture that "no turning back" moment. As a founding member of OpenAI and the former Director of AI at Tesla, Karpathy is one of the most respected architects of the modern AI era, lending his "boots-on-the-ground" observations a level of authority that few in the industry can match. He describes moving from 20% agent-assisted work to a staggering 80% in just a few weeks—essentially programming in English while managing the AI like a hasty but relentless junior dev. It's a fascinating look at how our roles are evolving from imperative "how-to" coding to declarative "what-if" strategy; we're trading the drudgery of syntax for the leverage of high-level code actions. While he warns of the "slopacolypse" and the subtle conceptual errors that still require a watchful eye in an IDE, the real takeaway is the massive expansion of what's now possible to build when stamina and "knowledge-skill issues" are no longer the primary bottlenecks.

Key Takeaways from Karpathy's Notes:

• The 80/20 Flip: Programming is becoming a "language-first" activity, where the human provides the intent and the agent handles the bulk of the implementation.
• The Stamina Advantage: AI agents don't get demoralized; their ability to relentlessly loop until they hit a success criterion provides a massive productivity multiplier.
• Junior Dev Pitfalls: Current models excel at syntax but struggle with "bloat" and making wrong assumptions—Karpathy emphasizes "watching them like a hawk" in a side-by-side IDE.
• Expansion Over Speed: The real value isn't just doing things faster; it's being able to tackle projects that previously weren't worth the time or were outside your skill set.
• The 2026 "Slopacolypse": As AI-generated content and code flood the web, the gap between "productivity theater" and actual, high-leverage engineering will widen significantly.